Generation of Face Privacy-Protected Images Based on the Diffusion Model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Growing concerns over personal data misuse with AI technologies necessitate advanced privacy solutions.
• Existing facial privacy methods often compromise visual quality, introduce distortions, or limit reusability.

Purpose of the Study:

• To introduce Diffusion Models for Face Privacy Protection (DIFP), a novel approach to enhance facial privacy.
• To address limitations of current methods by producing high-resolution, photorealistic, and reusable encrypted faces.

Main Methods:

• Utilized a conditionally controlled, reality-guided face generator for high-resolution encrypted face production.
• Employed a two-stage training strategy for identity and style guidance, enhanced by iterative latent variable improvement.
• Introduced diffusion model denoising for identity recovery and restoration of original facial data.

Main Results:

• Generated photorealistic encrypted faces that preserve naturalness and recoverability of original facial information.
• Achieved high success rates in evading face-recognition tools, demonstrating effective privacy protection.
• Enabled near-perfect restoration of occluded faces, showcasing the method's utility.

Conclusions:

• DIFP effectively protects facial privacy against AI-driven data misuse.
• The method balances robust privacy with the ability to recover original facial data when needed.
• DIFP represents a significant advancement in secure and reliable facial information management.